I build small research and developer tools in Python and JavaScript.
Paper Radar turns plain-English research interests into ranked paper reports and RSS feeds. It supports Jev and Claude backends, with a CLI, GitHub Action and an uploadable Claude skill. Python 3.11+, MIT, no runtime dependencies.
- Try the public examples · 繁體中文 — browse reports before installing, choose a research interest, and copy a ready-to-use prompt.
- Install in Claude · Download the upload ZIP — upload and enable skill 0.2.0, then describe an interest or paste arXiv/PubMed/DOI links. It retrieves public titles/abstracts and ranks in the current conversation; no CSV preparation, pip or separate model API key. Retrieval needs permitted network access or existing Claude web tools; normal Claude plan usage applies.
- Claude API quickstart · 繁體中文教學
- Ready-to-fork Claude example
- Read the public Jev feeds — no setup or key
- Try the fictional, offline demo:
python -m pip install paper-radar==0.2.0, thenpaper-radar demo
The upload bundle was tested in a Claude web conversation with Opus 5.5: a 44-paper fictional demo and a three-paper conversation report with HTML, CSV, JSON and RSS. These are functionality checks, not accuracy benchmarks. Scheduled CLI judging uses the separately billed Anthropic API. Claude scores are self-reported; the Jev benchmark does not establish Claude accuracy. Version 0.2.0 adds judging-request usage auditing, including repair and failed attempts, and reports unknown costs when usage or rates are missing.
s1kit provides confidence-based routing with Jev, local tests and threshold fitting. Check fitted thresholds on a separate labelled batch before relying on their accuracy.
app-xray inspects Electron applications and renders saved scan evidence. Installation currently uses the documented source build.
Setup questions and concrete experience reports are welcome in each project's issues or discussions.



